False tooth personalized generation system based on cooperation of multi-condition retrieval and lightweight fine tuning

Abstract

The invention relates to the technical field of computer aided design and manufacturing, and discloses a false tooth personalized generation system based on multi-condition retrieval and lightweight fine tuning collaboration, comprising a data preprocessing module used for establishing a standard coordinate system and unifying data; the offline library construction and training module and the online multi-source retrieval module are used for generating a query vector according to missing tooth environment characteristics by utilizing an aggregator network of contrast learning training and retrieving an optimal template point cloud; the personalized generation network module is used for fusing environmental constraints and template features through a trans-attention mechanism and predicting a displacement field to generate a rough denture model; and the function fine tuning module is used for carrying out iterative optimization by taking the vertex coordinates as variables based on occlusion and adjacent physical constraints. According to the method, the problems of fuzzy details and physical interference of the generated model are solved by utilizing anatomical priori of a real template and physical fine tuning during testing, and the form naturalness and clinical suitability of the false tooth are improved.

Description

Technical Field

[0001] This invention relates to the field of computer-aided design and manufacturing technology, specifically to a personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration. Background Technology

[0002] In the digital dental restoration process, the three-dimensional morphological design of dentures is a crucial step. Current automated design technologies mainly fall into two categories: deformation methods based on standard template libraries and generative methods based on deep learning.

[0003] Methods based on standard template libraries utilize pre-defined general tooth models and adapt them to edentulous areas using Laplacian deformation or freeform deformation algorithms. However, these methods are limited by the capacity of the template library; general standard templates struggle to cover the complex anatomical variations in the population. Furthermore, simple geometric deformation algorithms often result in distorted final models when faced with specific edentulous spaces due to significant differences between the initial template and the target morphology, failing to accurately reproduce the biological characteristics of natural teeth.

[0004] Deep learning-based generative methods learn statistical patterns from large amounts of case data through neural networks and directly output 3D models. While these methods improve the automation of the design process, the network models tend to generate averaged geometries, leading to the loss of key high-frequency anatomical details such as cusps and fissures, and resulting in overly smooth surfaces. Furthermore, data-driven generative networks lack explicit constraints on the physical environment; the generated models are only statistically sound, but in actual 3D space, they often cause penetrating interference with opposing teeth or poor contact with adjacent teeth, failing to meet clinical occlusion and proximal contact standards, thus requiring extensive manual adjustments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration. This system solves the technical problems of lost anatomical details on the model surface in existing automatic denture generation technologies, as well as the lack of physical space constraints leading to penetration interference or gap deviation between the generated dentures and surrounding teeth, which prevents them from directly meeting the requirements of clinical applications.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration, comprising: a data preprocessing module, used to receive the patient's oral cavity scan data, identify key anatomical feature regions to construct a standard coordinate system, and use a rigid body transformation matrix to unify the oral cavity scan data to the standard coordinate system;

[0007] The offline library construction and training module is used to build a partitioned index library containing the geometric features of teeth in historical cases, and to train an aggregator network based on a contrastive learning strategy. The aggregator network is used to map the multi-source environmental features of missing teeth into target query vectors.

[0008] The online multi-source retrieval module is used to extract local environmental data around the current patient's missing tooth position, generate the target query vector through the aggregator network, and retrieve the template tooth point cloud with the highest similarity to the target query vector in the partitioned index library;

[0009] A personalized generative network module is used to receive the template tooth point cloud and the target query vector, fuse the environmental constraints represented by the target query vector into the geometric features of the template tooth point cloud through a cross-attention mechanism, and predict the non-rigid displacement field to generate a rough denture model.

[0010] The function fine-tuning module is used to receive the rough denture model, and while keeping the neural network weight parameters locked, it uses the three-dimensional coordinates of the model vertices as optimization variables to iteratively optimize the rough denture model by minimizing the physical energy function containing occlusal constraints and adjacency constraints, and outputs the final denture model.

[0011] Preferably, the data preprocessing module constructs the standard coordinate system in the following manner:

[0012] The occlusal plane is obtained by fitting the set of cusps in the mandibular posterior tooth region, and the normal vector of the occlusal plane is defined as the Z-axis direction;

[0013] A midsagittal plane is constructed using the centroid of the central region of the incisor, with the normal vector of the midsagittal plane perpendicular to the Z-axis direction, and the normal vector of the midsagittal plane is defined as the X-axis direction.

[0014] According to the right-hand rule, the Y-axis direction is calculated using the cross product of the vectors in the Z-axis direction and the X-axis direction;

[0015] The vertical projection point of the central incisor's incisal point onto the occlusal plane is defined as the origin of the coordinate system.

[0016] Preferably, the offline library construction and training module trains the aggregator network in the following way:

[0017] For teeth known to exist in historical data, simulate their missing state, extract the corresponding symmetrical tooth features, adjacent tooth features, and opposing tooth features as environmental input, and construct the input tensor by combining the existence mask;

[0018] The input tensor is fed into the aggregator network to output the predicted query vector;

[0019] Construct a contrastive loss function, the objective of which is to maximize the cosine similarity between the predicted query vector and the actual tooth feature vector at that location, and minimize the similarity between the predicted query vector and the randomly sampled negative sample feature vector;

[0020] The parameters of the aggregator network are updated using the gradient descent algorithm.

[0021] Preferably, the online multi-source retrieval module performs the retrieval in the following specific manner:

[0022] The corresponding sub-database is determined based on the FDI number of the missing tooth position;

[0023] The point clouds of symmetrical teeth corresponding to the missing teeth of the patient are extracted and mirrored. The point clouds of adjacent teeth and opposing teeth are extracted, and the feature vectors of each are extracted using a geometric feature encoder.

[0024] A mask vector is generated based on the actual acquisition of each part of the point cloud. The feature vector and the mask vector are concatenated and then input into the aggregator network to obtain the target query vector.

[0025] Calculate the dot product between the target query vector and all template feature vectors in the partitioned sub-database, and select the point cloud corresponding to the template feature vector with the largest value as the template tooth point cloud.

[0026] Preferably, the cross-attention mechanism of the personalized generation network module is implemented as follows:

[0027] The local geometric features of the template tooth point cloud are mapped into a query matrix through a first projection matrix;

[0028] The target query vector is mapped into a key vector and a value vector through the second projection matrix and the third projection matrix, respectively, and the dimension is expanded through a broadcast mechanism.

[0029] Calculate the product of the query matrix and the transposed key vector, and then scale and normalize the product result using an exponential function to obtain the attention weights.

[0030] The value vector is weighted and summed using the attention weights, and the result is residually connected with the local geometric features of the original template tooth point cloud to obtain the fused feature matrix.

[0031] Preferably, the personalized generation network module further includes a decoder and a surface reconstruction unit;

[0032] The decoder employs a multilayer perceptron structure to process the fused feature matrix point by point, regress and predict the three-dimensional displacement of each vertex relative to its original position, and adds the vertex coordinates of the template tooth point cloud with the three-dimensional displacement to obtain a point cloud-form generative model.

[0033] The surface reconstruction unit uses the Poisson surface reconstruction algorithm to transform the point cloud-form generative model into a triangular mesh model with a closed manifold structure.

[0034] Preferably, the physical energy function constructed by the functional fine-tuning module is composed of a weighted sum of a biting loss term, an adjacency loss term, and a smoothing regularization term;

[0035] The function fine-tuning module uses the gradient descent algorithm to calculate the gradient of the physical energy function relative to the vertex coordinates, and updates the three-dimensional coordinates of the vertex according to the gradient until the physical energy function converges or reaches the preset number of iterations.

[0036] Preferably, the function fine-tuning module calculates the bite loss term in the following way:

[0037] Calculate the signed distance field of the opposing tooth surface model;

[0038] Query the distance value of each vertex of the rough denture model in the signed distance field;

[0039] Set a safe bite gap threshold, and use a modified linear unit function to filter out vertices with distance values ​​less than the safe bite gap threshold;

[0040] The sum of the squared distance deviations of the selected vertices is accumulated as the bite loss term to penalize areas where penetration or insufficient gap occurs.

[0041] Preferably, the function fine-tuning module calculates the adjacency loss term in the following way:

[0042] Identify the set of vertices located in the contact regions of the mesial and distal surfaces in the rough denture model;

[0043] Calculate the nearest Euclidean distance from each vertex in the vertex set to the nearest neighboring point cloud;

[0044] Set the lower and upper limits of the ideal contact distance range;

[0045] For distance deviations below the lower limit and distance deviations above the upper limit, squared penalties are calculated respectively, and all penalty values ​​are summed as the adjacency loss term.

[0046] Preferably, a personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration includes the following steps:

[0047] S1. Receive the patient's oral cavity scan data through the data preprocessing module, identify key anatomical feature regions to construct a standard coordinate system, and use a rigid body transformation matrix to unify the oral cavity scan data to the standard coordinate system.

[0048] S2. Construct a partitioned index library containing the geometric features of teeth in historical cases through an offline library construction and training module, and train an aggregator network based on a contrastive learning strategy. The aggregator network is used to map the environmental features of missing teeth into target query vectors.

[0049] S3. Extract local environmental data around the current patient's missing tooth position through the online multi-source retrieval module, generate the target query vector through the aggregator network, and retrieve the template tooth point cloud with the highest similarity to the target query vector in the partitioned index library;

[0050] S4. The template tooth point cloud and the target query vector are received through the personalized generation network module. The environmental constraints represented by the target query vector are fused into the geometric features of the template tooth point cloud using a cross-attention mechanism, and a non-rigid displacement field is predicted to generate a rough denture model.

[0051] S5. Receive the rough denture model through the function fine-tuning module, construct a physical energy function that includes occlusal non-penetration constraints and adjacent contact constraints, and generate the final denture model by iteratively optimizing the vertex coordinates of the rough denture model without updating the network parameters.

[0052] This invention provides a personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration. It has the following beneficial effects:

[0053] 1. This invention constructs a partitioned index library and trains an aggregator network. It uses natural tooth data from historical real cases as geometric priors. The aggregator network can retrieve the template with the highest matching degree based on the multi-source environmental features of the missing tooth position, providing an accurate initial topological structure for the subsequent generation process. This avoids the surface noise or structural distortion that may be generated by directly generating point clouds through neural networks, ensuring that the generated denture model conforms to normal physiological characteristics in terms of anatomical morphology.

[0054] 2. This invention utilizes a cross-attention mechanism to integrate patient-specific environmental features into standard template features, thereby guiding the prediction of non-rigid displacement fields. This mechanism enables the standard template to undergo local adaptive deformation based on the boundary morphology of the edentulous space while maintaining the integrity of the overall geometric structure. This solves the problem that general templates are difficult to accurately adapt to the specific dental arch spaces of different patients, and improves the geometric matching degree between the denture margin and adjacent and opposing teeth.

[0055] 3. After generating the network, this invention introduces a fine-tuning step based on physical constraints. By minimizing the energy function that includes occlusion and adjacency relationships, the grid vertex coordinates are directly optimized. This step is independent of the weight update of the neural network and can correct the penetration interference or poor contact between the model and the opposing teeth in real time for specific cases. This makes up for the small geometric errors that may exist in the prediction results of the neural network and ensures that the final output denture model can meet the occlusal gap and adjacency contact standards required for clinical application. Attached Figure Description

[0056] Figure 1 This is a system framework diagram of the present invention;

[0057] Figure 2 This is a flowchart of the method of the present invention.

[0058] The module includes: 10. Data preprocessing module; 20. Offline library construction and training module; 30. Online multi-source retrieval module; 40. Personalized network generation module; and 50. Function fine-tuning module. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example:

[0061] Please see the appendix Figure 1 - Appendix Figure 2 This invention provides a personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration. The system includes: a data preprocessing module 10, an offline library construction and training module 20, an online multi-source retrieval module 30, a personalized generation network module 40, and a function fine-tuning module 50.

[0062] The data preprocessing module 10 receives the patient's three-dimensional oral cavity scan data, including the maxillary model, mandibular model, and FDI number information identifying the missing teeth. The data preprocessing module 10 also establishes a standardized oral cavity coordinate system based on the three-dimensional oral cavity scan data and transforms the raw point cloud data to this unified spatial reference. This coordinate system is established based on anatomical feature points to ensure spatial consistency of data from different cases.

[0063] The offline library construction and training module 20 is used to process historical case data during the system preparation phase. This module is responsible for extracting geometric features from historical tooth data, constructing a feature vector index library for partitions, and training an aggregator network capable of mapping multi-source environmental information around missing teeth into a single query vector. The offline library construction and training module 20 provides the data retrieval foundation and feature mapping model for the online processing phase.

[0064] The online multi-source retrieval module 30 is connected to the data preprocessing module 10 and the offline library construction and training module 20. This module is used to extract the geometric features of the symmetrical teeth, adjacent teeth, and opposing teeth corresponding to the current patient's missing teeth, and to generate a comprehensive query vector using an aggregator network. The online multi-source retrieval module 30 further retrieves the template tooth data with the highest similarity from the feature vector index library based on this query vector, which serves as the geometric basis for subsequent generation.

[0065] The personalized generation network module 40 is connected to the online multi-source retrieval module 30 to receive the retrieved template tooth data and query vectors. This module uses a cross-attention mechanism to calculate the correlation between the template tooth point cloud and the current oral environment features, and adjusts the morphology of the template tooth through a non-rigid deformation algorithm to adapt it to the current patient's edentulous gap and adjacency relationship, outputting a preliminary adapted denture point cloud model.

[0066] The functional fine-tuning module 50 is connected to the personalized generation network module 40 and is used to optimize the physical constraints of the initially adapted denture point cloud model. This module constructs an objective function that includes occlusal non-penetration terms, adjacent contact terms, and smoothing regularization terms, and uses the gradient descent algorithm to iteratively adjust the position of the denture surface vertices, outputting a final three-dimensional denture model that meets clinical space requirements.

[0067] In the overall workflow of the system, the data preprocessing module 10 first cleans and standardizes the input maxillary and mandibular scan data to eliminate differences in scanning posture. The processed and standardized data is then transmitted to the online multi-source retrieval module 30.

[0068] The online multi-source retrieval module 30 extracts environmental features from standardized data based on missing tooth position information, and calls the aggregator parameters and feature library indexes pre-stored in the offline library construction and training module 20 to quickly locate and output the reference template that best matches the current anatomical environment.

[0069] The reference template and its corresponding environmental feature vector are then input into the personalized generation network module 40. The personalized generation network module 40 performs geometric deformation operations, transforming the standardized reference template into an intermediate model with individualized morphological features.

[0070] Finally, the intermediate model is fed into the functional fine-tuning module 50. Based on preset biomechanical constraints, the functional fine-tuning module 50 calculates the spatial distance relationship between the model and the opposing and adjacent teeth, and makes minor adjustments to the vertex displacement of the model surface until the preset contact and occlusion standards are met, thus completing the denture generation process.

[0071] The data preprocessing module 10 is mainly used to perform format conversion and spatial standardization on the input raw oral cavity scan data, providing a unified geometric benchmark for subsequent feature extraction and model generation. This processing specifically includes the following steps:

[0072] S101, Data Reception and Format Conversion. The data preprocessing module 10 receives the patient's maxillary and mandibular 3D models. The 3D models are typically stored in a standard triangular mesh format, along with FDI number information identifying missing teeth and other tooth positions. For mesh-formatted data, the module converts it into point cloud data format through vertex sampling or mesh vertex extraction.

[0073] S102, Key Feature Region Identification. To establish a coordinate system conforming to oral anatomy, it is necessary to identify feature regions in the point cloud data for fitting a reference plane. Using FDI numbering information, a feature point detection algorithm based on discrete average curvature or a method based on local height extrema is employed to select the set of cusps in the posterior tooth region from the mandibular point cloud; simultaneously, a set of center points in the central incisor region is selected from the maxillary or mandibular point cloud.

[0074] S103, Occlusal plane fitting. The data preprocessing module 10 uses a robust estimation method to perform plane fitting on the feature point set to obtain the occlusal plane. In this embodiment, the normal vector of the plane is calculated using a random sampling consensus algorithm or principal component analysis, and this vector corresponds to the Z-axis direction in the standard coordinate system;

[0075] S104, Midsagittal Plane Construction. The midsagittal plane is defined as a bilaterally symmetrical reference plane. This plane must satisfy two geometric constraints: it passes through the centroid of the incisor center region, and its normal vector is perpendicular to the normal vector of the occlusal plane. The normal vector of this plane is solved using the constrained least squares method, corresponding to the X-axis direction in the standard coordinate system;

[0076] S105, Coordinate System Construction and Transformation Matrix Calculation. The origin of the standard coordinate system is defined as the intersection of the midsagittal plane and the dental arch curve, or simply as the perpendicular projection of the central incisor's incisal point onto the occlusal plane. Based on the determined Z-axis and X-axis direction vectors, the Y-axis direction vector is calculated using the right-hand rule. This constructs the rigid body transformation matrix from the original coordinate system to the standard coordinate system. This transformation matrix includes rotation matrices and translation vectors, used to map points in the original coordinate system to the standard coordinate system.

[0077] S106, Data Standardization Transformation. Using the calculated rigid body transformation matrix, the coordinates of each point in the original point cloud set are transformed. After the above steps, the oral cavity scan data of all patients are unified to the same anatomical spatial reference, eliminating spatial differences caused by different scanning device postures or patient head positions.

[0078] The offline library construction and training module 20 is mainly used to build a geometry-based retrieval index library and train a feature aggregation network that can fuse multi-source anatomical environment information. The operation of this module specifically includes the following steps:

[0079] S201, Individual Tooth Feature Extraction. The offline library construction and training module 20 processes each case in the historical case database. For each tooth in the case, it acquires the corresponding standardized 3D point cloud data. The module calls a pre-built geometric feature encoder to encode the point cloud data. In this embodiment, the geometric feature encoder adopts a network architecture that can directly process unordered point clouds and extract global features with rotation and translation invariance;

[0080] S202, Construction of the Partition Feature Index Library. To improve the efficiency and accuracy of subsequent retrieval, this module constructs a feature vector index library. This index library adopts a hierarchical storage strategy, dividing the feature vectors into multiple independent sub-libraries based on the FDI number of the teeth. Within each sub-library, an index structure is constructed to support fast similarity retrieval, such as an index based on the inner product;

[0081] S203, Multi-source Environmental Feature Construction. To train the aggregator network, model input data needs to be constructed. For any known tooth in the historical data, a simulated environmental state when it is missing is constructed. The module extracts four types of environmental constraint data corresponding to the target tooth: symmetrical tooth features, left neighbor tooth features, right neighbor tooth features, and opposing tooth features. The symmetrical tooth features are obtained by acquiring the point cloud of the corresponding tooth in the opposite quadrant of the target tooth, mirroring it along the midsagittal plane, and then extracting it. Simultaneously, an existence mask vector is constructed to indicate the presence or absence of features in the above four dimensions. The feature vectors and the mask vector are concatenated to form the input tensor of the aggregator network.

[0082] S204, Aggregator Network Construction. Construct a neural network-based aggregator to map high-dimensional concatenated inputs into a single query vector. This network consists of an input layer, several hidden layers, and an output layer. The output query vector lies on a unit hypersphere, with dimensions consistent with the dimension of a single tooth feature vector.

[0083] S205, Contrastive Learning Training. The aggregator network is trained under supervision using a contrastive loss function. The training objective is to improve the query vectors generated by the aggregator. Compared with the original tooth feature vector at that location (Positive samples) should be as similar as possible, while being as different as possible from other randomly sampled tooth feature vectors (negative samples) in the database. Define the loss function. as follows:

[0084] ;

[0085] Among them, symbols The dot product operation represents vectors, with the symbol... This represents the L2 norm of the vector. During training, the weight parameters of the aggregator network are updated using the gradient descent algorithm until the loss function converges. After training, the aggregator network can infer the geometric features that an ideal missing tooth should possess based on incomplete oral environment information.

[0086] The online multi-source retrieval module 30 is primarily responsible for generating a query vector that represents the ideal morphology of the target tooth in real time based on the current position of the patient's missing teeth and the surrounding anatomical environment, and retrieving the best matching template from the offline feature library. The specific steps involved in the operation of this module are as follows:

[0087] S301, Extraction of missing tooth location environmental data. The online multi-source retrieval module 30 first determines the missing tooth location identifier to be generated based on the input FDI number. Based on this identifier, the module extracts local point cloud data related to the missing tooth location space from standardized maxillary or mandibular point cloud data, including: symmetrical tooth point cloud, adjacent tooth point cloud, and opposing tooth point cloud;

[0088] S302, Real-time Feature Encoding. Utilizing a geometric feature encoder that shares parameters with the offline library construction phase, features are extracted from the local point clouds extracted in step S301, transforming the incomplete oral cavity geometry information of the current patient into a vector representation in a high-dimensional feature space;

[0089] S303, Target Query Vector Generation. This module constructs the feature tensor and mask vector of the current environment. The mask vector is determined based on the data existence state in S301. The extracted feature vector and mask vector are concatenated and input into a pre-trained aggregator network. The aggregator network performs weighted fusion of the input features and outputs a target query vector.

[0090] S304, Similarity-based full-database retrieval. The online multi-source retrieval module 30 locates the corresponding sub-database in the feature index based on the missing tooth position identifier. Within the sub-database, it calculates the similarity between the target query vector and all template feature vectors in the database. Based on the calculation results, the module returns the point cloud data of the template tooth with the highest similarity and its corresponding original feature vector, which serves as the geometric initialization basis for subsequent generation steps.

[0091] The personalized generative network module 40 is primarily used to convert standardized template tooth point clouds into personalized models adapted to the patient's specific anatomical environment. This process is achieved by predicting non-rigid deformation fields using deep neural networks, and specifically includes the following steps:

[0092] S401, Generate Network Input Construction. The module receives the retrieved template tooth point cloud and the target query vector. The target query vector encodes the local anatomical environment features of the missing tooth location. The module first maps the 3D coordinates of each vertex in the template point cloud to high-dimensional local geometric features, forming a template feature matrix;

[0093] S402, Cross-Attention Feature Fusion. To integrate the patient's oral environment constraints into the geometric features of the standard template, the module employs a cross-attention mechanism for feature fusion. In this mechanism, the local geometric features of the template serve as queries, and the environmental features serve as keys and values.

[0094] Calculate the fused feature matrix The process can be summarized as follows:

[0095] ;

[0096] in:

[0097] The query projection matrix for template features;

[0098] and These are the key projection vector and value projection vector of the environmental features, respectively;

[0099] This is the original template feature matrix, which is used here as a residual connection term;

[0100] The dimension of the feature channel serves as a scaling factor.

[0101] superscript This represents the matrix transpose operation;

[0102] This represents the normalized exponential function. During the calculation, the environmental feature vector is expanded to the same dimension as the template points through a broadcast mechanism;

[0103] S403, Non-rigid Displacement Field Prediction. A feature matrix incorporating environmental context information is input into the decoder network. The decoder network uses a multilayer perceptron to process each feature point in the feature matrix point by point, directly regressing to predict the 3D displacement of each vertex relative to its original position. This step utilizes the non-linear fitting capability of deep neural networks to cause morphological distortion and scaling of the standard template while maintaining the overall topological structure, in order to initially adapt to the spatial size and edge morphology of the missing tooth location.

[0104] S404, 3D Surface Reconstruction. Since the data output by the neural network is in the form of discrete point clouds, in order to meet the needs of subsequent physical fine-tuning and clinical manufacturing, the module uses a surface reconstruction algorithm to process the generated point clouds and obtain a triangular mesh model with a closed manifold structure.

[0105] The functional fine-tuning module 50 is primarily used to address potential biases in the clinical functionality of the geometric model generated by the deep neural network. This module directly uses the coordinates of the mesh vertices as optimization variables, achieving model fine-tuning by minimizing the physical constraint energy function. The process specifically includes the following steps:

[0106] S501, Optimize Environment Construction and Variable Definition. The functional fine-tuning module 50 receives the reconstructed denture mesh model from the personalized generation network module 40. Simultaneously, it acquires the surface model of the opposing tooth and the point cloud data of adjacent teeth as environmental constraint boundaries. The optimization variable is defined as the set of displacements of all vertices;

[0107] S502, Occlusal without Penetration Constraint Calculation. To prevent interference between the generated denture and the opposing teeth and to reserve an appropriate occlusal gap, the module constructs an occlusal loss function. This function queries the signed distance field of the tooth surface model and penalizes vertices that intrude into the safe clearance threshold.

[0108] ;

[0109] in:

[0110] This represents the total number of vertices in the denture mesh model.

[0111] This is the preset safe interlocking clearance threshold;

[0112] Let be a function representing a vertex. The signed distance value to the surface of the opposing tooth;

[0113] To correct the linear unit function, used to filter out legal vertices whose distance is greater than the safety threshold;

[0114] S503, Adjacent Contact Zone Constraint Calculation. To ensure a tight contact relationship between the denture and adjacent teeth, this module constructs an adjacency loss function. This function identifies the set of vertices located in the contact region of the denture mesh. This forces it to maintain its distance from adjacent tooth point clouds within the ideal range. Inside:

[0115] ;

[0116] in, Represents the vertex of the contact area Euclidean distance to the nearest neighbor in the adjacent point cloud. and These are the lower and upper limits of the contact distance, respectively.

[0117] S504, Surface Smoothing Regularization. To prevent mesh surface distortion during fine-tuning, the module introduces a smoothing regularization term to ensure that vertices maintain a relatively smooth positional relationship with their neighbors;

[0118] S505, Iterative Optimization Solution. The functional fine-tuning module 50 constructs the overall objective function, summing the three constraints mentioned above with weights. This module uses gradient descent to iteratively update the optimization variables until the overall loss function converges. After optimization, it outputs the final 3D model of the denture that conforms to the physical functional constraints. This process is completed in real-time during the testing phase, without relying on large-scale datasets for retraining.

[0119] Please see the appendix Figure 2 This invention provides a method for personalized denture generation based on multi-condition retrieval and lightweight fine-tuning collaboration, comprising the following steps:

[0120] S1. Obtain the patient's oral cavity three-dimensional scan data and missing tooth position information, construct a standardized oral cavity coordinate system, and transform the oral cavity three-dimensional scan data to the standardized oral cavity coordinate system;

[0121] S2. Construct an offline feature index library and train a multi-source feature aggregation network. The offline feature index library stores tooth templates of historical cases and their corresponding geometric feature vectors. The multi-source feature aggregation network is used to map the features of symmetrical teeth, adjacent teeth and opposing teeth into query vectors.

[0122] S3. Extract the multi-source environmental features corresponding to the current patient's missing tooth position, generate the target query vector using the multi-source feature aggregation network, and retrieve the reference template with the highest similarity to the target query vector from the offline feature index library.

[0123] S4. Input the reference template and target query vector into the preset generation network, use the cross-attention mechanism to calculate the association between the reference template and multi-source environmental features, and perform non-rigid deformation on the reference template to generate the initial denture point cloud.

[0124] S5. Construct an objective function that includes occlusal constraints and adjacency constraints, and use the gradient descent algorithm to iteratively optimize the vertex positions of the initial denture point cloud to obtain a three-dimensional denture model that meets clinical space requirements.

Claims

1. A personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration, characterized in that, include: The data preprocessing module (10) is used to receive the patient's oral cavity scan data, identify key anatomical feature regions to construct a standard coordinate system, and use a rigid body transformation matrix to unify the oral cavity scan data to the standard coordinate system. The offline library construction and training module (20) is used to construct a partitioned index library containing the geometric features of teeth in historical cases, and to train an aggregator network based on a contrastive learning strategy. The aggregator network is used to map the multi-source environmental features of missing teeth into target query vectors. The online multi-source retrieval module (30) is used to extract local environmental data around the missing tooth position of the current patient, generate the target query vector through the aggregator network, and retrieve the template tooth point cloud with the highest similarity to the target query vector in the partitioned index library; The personalized generation network module (40) is used to receive the template tooth point cloud and the target query vector, fuse the environmental constraints represented by the target query vector into the geometric features of the template tooth point cloud through a cross-attention mechanism, and predict the non-rigid displacement field to generate a rough denture model. The function fine-tuning module (50) is used to receive the rough denture model, and while keeping the neural network weight parameters locked, it uses the three-dimensional coordinates of the model vertex as the optimization variable, and iteratively optimizes the rough denture model by minimizing the physical energy function containing occlusal constraints and adjacency constraints, and outputs the final denture model.

2. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 1, characterized in that, The specific method by which the data preprocessing module (10) constructs the standard coordinate system is as follows: The occlusal plane is obtained by fitting the set of cusps in the mandibular posterior tooth region, and the normal vector of the occlusal plane is defined as the Z-axis direction; A midsagittal plane is constructed using the centroid of the central region of the incisor, with the normal vector of the midsagittal plane perpendicular to the Z-axis direction, and the normal vector of the midsagittal plane is defined as the X-axis direction. According to the right-hand rule, the Y-axis direction is calculated using the cross product of the vectors in the Z-axis direction and the X-axis direction; The vertical projection point of the central incisor's incisal point onto the occlusal plane is defined as the origin of the coordinate system.

3. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 1, characterized in that, The offline library construction and training module (20) trains the aggregator network in the following way: For teeth known to exist in historical data, simulate their missing state, extract the corresponding symmetrical tooth features, adjacent tooth features, and opposing tooth features as environmental input, and construct the input tensor by combining the existence mask; The input tensor is fed into the aggregator network to output the predicted query vector; Construct a contrastive loss function, the objective of which is to maximize the cosine similarity between the predicted query vector and the actual tooth feature vector at that location, and minimize the similarity between the predicted query vector and the randomly sampled negative sample feature vector. The parameters of the aggregator network are updated using the gradient descent algorithm.

4. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 1, characterized in that, The online multi-source retrieval module (30) performs the retrieval in the following specific way: The corresponding sub-database is determined based on the FDI number of the missing tooth position; The point clouds of symmetrical teeth corresponding to the missing teeth of the patient are extracted and mirrored. The point clouds of adjacent teeth and opposing teeth are extracted, and the feature vectors of each are extracted using a geometric feature encoder. A mask vector is generated based on the actual acquisition of each part of the point cloud. The feature vector and the mask vector are concatenated and then input into the aggregator network to obtain the target query vector. Calculate the dot product between the target query vector and all template feature vectors in the partitioned sub-database, and select the point cloud corresponding to the template feature vector with the largest value as the template tooth point cloud.

5. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 1, characterized in that, The specific implementation of the cross-attention mechanism of the personalized generation network module (40) is as follows: The local geometric features of the template tooth point cloud are mapped into a query matrix through a first projection matrix; The target query vector is mapped into a key vector and a value vector through the second projection matrix and the third projection matrix, respectively, and the dimension is expanded through a broadcast mechanism. Calculate the product of the query matrix and the transposed key vector, and then scale and normalize the product result using an exponential function to obtain the attention weights. The value vector is weighted and summed using the attention weights, and the result is residually connected with the local geometric features of the original template tooth point cloud to obtain the fused feature matrix.

6. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 1, characterized in that, The personalized generation network module (40) also includes a decoder and a surface reconstruction unit; The decoder employs a multilayer perceptron structure to process the fused feature matrix point by point, regress and predict the three-dimensional displacement of each vertex relative to its original position, and adds the vertex coordinates of the template tooth point cloud with the three-dimensional displacement to obtain a point cloud-form generative model. The surface reconstruction unit uses the Poisson surface reconstruction algorithm to transform the point cloud-form generative model into a triangular mesh model with a closed manifold structure.

7. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 1, characterized in that, The physical energy function constructed by the functional fine-tuning module (50) is composed of a weighted sum of the biting loss term, the adjacency loss term, and the smoothing regularization term; The function fine-tuning module uses the gradient descent algorithm to calculate the gradient of the physical energy function relative to the vertex coordinates, and updates the three-dimensional coordinates of the vertex according to the gradient until the physical energy function converges or reaches the preset number of iterations.

8. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 7, characterized in that, The function fine-tuning module (50) calculates the bite loss term in the following way: Calculate the signed distance field of the opposing tooth surface model; Query the distance value of each vertex of the rough denture model in the signed distance field; Set a safe bite gap threshold, and use a modified linear unit function to filter out vertices with distance values ​​less than the safe bite gap threshold; The sum of the squared distance deviations of the selected vertices is accumulated as the bite loss term to penalize areas where penetration or insufficient gap occurs.

9. The personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration according to claim 7, characterized in that, The function fine-tuning module (50) calculates the adjacency loss term in the following way: Identify the set of vertices located in the contact regions of the mesial and distal surfaces in the rough denture model; Calculate the nearest Euclidean distance from each vertex in the vertex set to the nearest neighboring point cloud; Set the lower and upper limits of the ideal contact distance range; For distance deviations below the lower limit and distance deviations above the upper limit, squared penalties are calculated respectively, and all penalty values ​​are summed as the adjacency loss term.

10. A personalized denture generation system based on multi-condition retrieval and lightweight fine-tuning collaboration, comprising the steps described in any one of claims 1-9: S1. Receive the patient's oral cavity scan data through the data preprocessing module (10), identify key anatomical feature areas to construct a standard coordinate system, and use the rigid body transformation matrix to unify the oral cavity scan data to the standard coordinate system; S2. Construct a partitioned index library containing the geometric features of teeth in historical cases through the offline library construction and training module (20), and train an aggregator network based on the contrastive learning strategy. The aggregator network is used to map the environmental features of missing teeth into target query vectors. S3. Extract local environmental data around the current patient's missing tooth position through the online multi-source retrieval module (30), generate the target query vector through the aggregator network, and retrieve the template tooth point cloud with the highest similarity to the target query vector in the partitioned index library; S4. The template tooth point cloud and the target query vector are received by the personalized generation network module (40). The environmental constraints represented by the target query vector are fused into the geometric features of the template tooth point cloud by the cross-attention mechanism, and the non-rigid displacement field is predicted to generate a rough denture model. S5. The rough denture model is received through the function fine-tuning module (50), and a physical energy function containing occlusal non-penetration constraints and adjacent contact constraints is constructed. Without updating the network parameters, the vertex coordinates of the rough denture model are iteratively optimized to generate the final denture model.

Images